Cross-Source Verification

The system requires a minimum number of independent sources to agree on a factual claim before it is included in the generated answer. This threshold is configurable:

• Strict mode: Requires 3+ corroborating documents for high-stakes domains like medicine or law
• Standard mode: Requires 2+ sources for general knowledge queries
• Single-source fallback: Allowed only when explicitly cited with a low-confidence flag

The consensus mechanism treats each source as a voter, and only claims that cross a predefined agreement threshold survive filtering.

Verification requires that corroborating documents originate from genuinely independent origins, not syndicated or derivative content. The system checks:

• URL domain diversity: Documents must come from different root domains
• Attribution chains: Sources that cite the same original report are deduplicated and treated as a single origin
• Temporal independence: A source published after another that merely restates it is flagged as non-independent

This prevents circular confirmation where one original claim proliferates across mirrors and appears as multiple confirmations.

When sources disagree, the system does not simply pick a majority winner. Instead, it executes a structured conflict resolution pipeline:

• Authority weighting: Sources with higher Source Reliability Scores are given more evidentiary weight
• Recency analysis: For time-sensitive facts, newer sources may override older ones if a clear temporal trend exists
• Granular claim decomposition: A complex statement is broken into atomic sub-claims, each verified independently
• Unresolved conflict handling: If contradiction persists, the system surfaces the disagreement to the user with inline citations for both positions rather than fabricating a resolution

Before cross-referencing, generated text is decomposed into atomic factual assertions—single, verifiable claims that cannot be further subdivided. Each atom is then:

• Normalized: Entity mentions are disambiguated and dates standardized
• Embedded: Converted to a dense vector for semantic comparison against source chunks
• Aligned: Matched to specific spans in retrieved documents using Natural Language Inference (NLI) to determine entailment, contradiction, or neutrality

This granular approach prevents a document that supports 90% of a paragraph from being counted as full corroboration when it contradicts a critical detail.

The system outputs a calibrated confidence score that directly correlates with the degree of cross-source agreement:

• High confidence (0.9+): Claim verified by 3+ independent, high-authority sources with zero contradictions
• Medium confidence (0.7–0.89): Claim verified by 2 sources or has minor temporal discrepancies
• Low confidence (<0.7): Single-source claim or unresolved contradiction; surfaced with explicit caveats

This score is not a static model probability but a dynamic metric computed from the retrieval graph, enabling downstream systems to make risk-aware decisions about whether to display, suppress, or escalate information.

Cross-source verification incorporates temporal context to prevent outdated consensus from overriding new facts. The system:

• Timestamps every source at ingestion and tracks content freshness
• Detects consensus drift: When newer sources systematically diverge from older ones, the system flags a potential factual update rather than treating the old consensus as verified
• Applies decay functions: Older corroborations lose weight over time for rapidly evolving topics

This prevents scenarios where a widely-cited but obsolete claim continues to pass verification simply because many sources once repeated it.